Content Quality: Well-structured News piece (Overview, What We Know, The License, Cyber-Capability Controls, Two Benchmark Scores, What We Don't Know). Body is 1,098 words with markup and link targets stripped (1,104 by the pipeline counter, within the 400-1200 News range). Title is 142 characters, under the 150-character limit mentioned for review; long but accurate and every element is supported by the license text. The licence section reproduces the three numbered conditions (sections 2, 3, 4) with exactly what each measures and no legal characterisation, and the article never calls the release open source or not open source.
Source Verification: All six snapshots (source-0..5.html.gz, all HTTP 200, suspicious_patterns null for every source) were decompressed and read; the license and model card were additionally pulled raw from huggingface.co (raw/main/LICENSE, raw/main/README.md, /api/models) and the Moonshot blog was re-fetched from kimi.ai to cross-check. (1) source-0 www.kimi.com blog 'Kimi K3.1 Pro Tech Blog': confirms verbatim the quotes 'the first model in our new K3.1 series', 'The full model weights will be released by Nov 4.', 'identified and removed cyber-related data from both training stages', and 'remains behind that of frontier proprietary models that have been publicly available for at least three months.' (article ends the quote at 'months.' which is the source's own text); availability on Kimi.com, Kimi Work, Kimi Code, Kimi API and the Highspeed variant confirmed; xdev-bench-hard 90.5% (checkpoint without cyber-data removal at mid-training and RL environment filtering) and 81.0% (final model), pass@3 over a random subset of runs, Moonshot's in-house benchmark: confirmed and correctly labelled as Moonshot's own; Terminal-Bench 4.0 65.7% vs GPT-6 Astra 58.2% and Claude Opus 5.5 66.4%: confirmed. (2) source-1 Hugging Face model card: MoE, 2.8T total, 104B activated, context 1048576, MXFP4 weights / MXFP8 activations, modality Text, Image, vLLM/SGLang/TokenSpeed recommended, platform.kimi.ai with model name kimi-k3.1-pro, reasoning_effort low/high/max default max, released under the Kimi K3.1 Pro License: all confirmed. HF API: repo created 2026-11-02, last modified 2026-11-05, not gated; weights are public (112 files, 96 safetensors shards). (3) source-2 LICENSE (raw text): titled 'Kimi K3.1-Pro License', 'Copyright (c) 2026 Moonshot AI'. Section 2 defines Model as a Service as giving a third party access to language model inference or fine-tuning (e.g. via API) with meaningful control over inputs, parameters or training data, excluding (a) end-user products with embedded capabilities and (b) mere relaying; condition: if the Licensee or any affiliates operates a MaaS business and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or equivalent) in total over any consecutive 12 months, a separate agreement with Moonshot AI is needed before using the Software or derivatives for any commercial purpose. So the $20M figure measures aggregate revenue (licensee plus affiliates, not MaaS revenue alone) over any rolling 12 months; the article states this accurately. Section 3: >100M MAU or >20M USD monthly revenue triggers display of 'Kimi K3.1-Pro' on the UI; the article's quote 'must be prominently displayed on the user interface' is verbatim (original reads 'must be prominently displayed on the
user interface'). Section 4 exemptions (internal use; use through Moonshot's official products or certified inference partners): confirmed. (4) source-3 The Decoder, Oct 23, 2026: 55 points on the AA Intelligence Index, +11 over K3, fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56), 'according to Artificial Analysis'; the sentence 'would become the highest-scoring open-weights model available' is verbatim; pricing $2/$10/$0.20 down from $3/$15/$0.30 and 33 percent cut; 90.5/81.0 percent and DeepSeek-V4.1-Flash 85.7 percent on xdev-bench-hard (Moonshot's table); Terminal-Bench 4.0 as stated. The Decoder does not state an index version. (5) source-4 TechNode, Nov 5, 2026: weights on Hugging Face, flagship multimodal model, 2.8T / ~104B activated / ~1M context, follows earlier availability via apps and API, Huawei inference support via vLLM Ascend and SGLang for Atlas A3 and Ascend 950: confirmed. (6) source-5 Artificial Analysis model page, snapshot of Nov 5, 2026: score 46 on the Intelligence Index, median 37, Intelligence Index v5.0.1 with nine named evaluations, header 'Kimi - Proprietary model - Released October 2026', FAQ 'No, Kimi K3.1 Pro is proprietary. The model weights are not publicly available.', $2.00 in / $10.00 out, filter text 'Indicates whether the model weights are available': all confirmed. The embedded page data additionally carries isOpenWeights=false and lists the Hugging Face repository and the Kimi license as the weights/license URLs, which the article's hedged statement ('the page does not say so') does not contradict. Discrepancy 55 (Decoder, Oct 23, citing AA) vs 46 (AA page, Nov 5, v5.0.1): both verified from raw pages; the article reports the gap and explicitly declines to explain it. Independent-source count: this is a News article, so the Analysis independence rule does not apply; independent of the subject are The Decoder, TechNode and Artificial Analysis (3 of 6); Moonshot blog, model card and LICENSE are the subject's own pages. Allowlist: the-decoder.com, technode.com and artificialanalysis.ai are already allowlisted; kimi.com was not. Verification that kimi.com is Moonshot AI's official site: Moonshot's own GitHub repo MoonshotAI/Kimi-K2 README and Moonshot's Hugging Face moonshotai/Kimi-K2-Instruct README both link https://www.kimi.com as the official Kimi chat; the Hugging Face K3.1 Pro model card links the same blog at www.kimi.ai/blog/kimi-k3-1-pro (Moonshot's other domain, linked from www.moonshot.ai), which serves the same text including all four quoted passages and the 90.5/81.0 figures. Confirmed, so kimi.com is added to config/source_allowlist.txt with justification, committed with this review. Pattern scanner: suspicious_patterns is null for all six sources; no injection text found in any snapshot, no override was needed.
Factual Accuracy: Every number, quote, version and date checked against the snapshots and raw pages and matched, with one exception: the 'What We Don't Know' bullet 'Whether any independent party has reproduced Moonshot's coding, finance or cyber figures. The sources reviewed did not report one.' is incomplete as to coding and agentic-work figures. The Artificial Analysis page (source-5), whose evaluations the site describes as 'measured independently by Artificial Analysis', lists Terminal-Bench 4.0 among the Intelligence Index evaluations, and its embedded page data gives Kimi K3.1 Pro a Terminal-Bench 4.0 score of 0.6717 (about 67.2%, versus the 65.7% Moonshot reports), and a GDPval figure; The Decoder separately reports a GDPval-AA score of 1,846 Elo 'according to Artificial Analysis'. These are independent measurements, not necessarily under Moonshot's conditions, but they are reports a reader would expect to be weighed. The statement is accurate for the cyber figures (xdev-bench-hard has no independent replication in any cited source), and the cyber-section sentence 'the Herald found no independent replication of them' is accurate. The Terminal-Bench paragraph correctly labels 65.7% as vendor-reported. Handled through a corrections record (clarification of one subordinate bullet, not the headline, summary or lead). Weights release date: the article does not assert a date; the blog says 'by Nov 4', TechNode is dated Nov 5, the Hugging Face repo was created Nov 2 and was public with 96 shards on Nov 5, none of which contradict the article. Pricing statements name their scope and date (The Decoder, Oct 23; Artificial Analysis, read Nov 5) and carry the 'may differ' caveat.
Overall Assessment: Accurate, carefully hedged coverage of the Kimi K3.1 Pro weights release; the licence terms, cyber-data figures and the 55 versus 46 index discrepancy are all verified from raw pages and presented without speculation. One subordinate claim about the absence of independent measurements is incomplete and is covered by a corrections record, so the verdict is APPROVE_WITH_CORRECTIONS. Featured candidacy: 1 - solid, well-sourced coverage of an open-weight release and its licence terms for a defined developer and AI-industry audience, not industry-changing